Confounding and Causation: Why a Real Association Can Still Lie to You

Two things can move together for years without one causing the other, and the usual reason is a third thing quietly steering both. That third thing is a confounder, and it is the most common reason an honest, well-measured association turns out to be false as a cause. When a study reports that people who do X have more of outcome Y, the disciplined first question is not "how strong is the link" but "what else differs between the people who do X and the people who do not." Good studies answer that on purpose, through randomization, adjustment, and design choices made before any data arrive. This piece is educational and not medical advice; for decisions about your own care, talk with your own clinician.

I work on both sides of this problem. My doctoral research at the Lund University Diabetes Centre is on the genetics of type 2 diabetes, where teasing a causal variant out of a sea of correlated markers is the whole job, so confounding is not an exotic footnote to me.

What is a confounder, in one sentence?

Here is the short, quotable version. A confounder is a factor associated with the suspected cause, which independently affects the outcome, and is not simply a step on the causal path between them. Ignored, it can manufacture an association out of nothing, hide a real one, or make a true effect look larger or smaller than it is. It is what lets correlation pose as cause.

A clean everyday example: ice cream and drowning

Across a year, ice cream sales and the number of drownings rise and fall together, yet nobody sensible concludes that ice cream pulls people under the water. The reason both numbers move is summer. Hot weather drives ice cream sales up, and the same heat sends more people swimming, which raises drownings. Temperature is the confounder. It is associated with the suspected cause (ice cream), it independently affects the outcome (more swimming, more drownings), and it sits off to the side rather than on a causal chain from cone to canal. Compare hot days only against other hot days, and the ice cream signal evaporates. The link was never about ice cream. It was about the weather, wearing an ice cream costume. Real confounders in health research are rarely this easy to spot, and they tangle into almost every comparison we care about.

Why this trips up serious research too

The dangerous confounders in clinical work hide inside the reasons people end up in one group rather than another. Picture an observational study finding that patients on a particular therapy do better than those not on it. The people who get offered, accept, and stay on a treatment are often those more engaged with their health, who can afford to fill prescriptions, who started less ill. Each of those traits independently improves outcomes, so the raw comparison cannot tell you whether the treatment helped, did nothing, or mildly harmed. The groups differed before it was ever given. This is not a story about careless people. The question is genuinely hard, because you cannot observe the world in which the same patient skipped the drug.

A sharper version, called confounding by indication, is where the very reason a treatment was chosen also predicts the outcome. The sickest patients get the strongest drugs, so those drugs can look dangerous in raw data, not because they harm but because they went to people already in trouble. The apparent direction can flip entirely.

How do good studies guard against confounding?

There are three broad defenses, and the best studies layer them.

Randomization, the strongest tool we have

The most powerful protection is to assign the exposure by chance. In a randomized controlled trial, a coin flip decides who receives the intervention. Because assignment is random, the two groups end up balanced on average across every confounder, including the ones nobody thought to measure and the ones nobody has even named. Adjustment can only correct for confounders you measured; randomization handles the unknown ones too, by construction.

I have lived inside this design. With EASY Diabetes, an AI clinical decision-support system for type 2 diabetes that I co-developed, we ran a multi-clinic randomized controlled trial, EASY-1, comparing the system against standard care. We randomized rather than comparing volunteer clinics with the rest for the reason described here: such clinics differ in funding, culture, and baseline performance.

Adjustment, useful but only as good as your list

When randomization is impossible or unethical, researchers measure the likely confounders and account for them statistically, through methods like regression, stratification, or matching. Done well, this removes much of the bias. The limit deserves stating plainly: adjustment can only neutralize confounders you thought of and measured well. An unmeasured one, or one measured crudely, slips straight through and leaves residual bias behind, sometimes under a reassuring show of rigor. So when a study leans entirely on adjustment, ask which confounders it could not measure, and which way those would push the result.

Design, the defenses you build before collecting data

Some of the best protection happens in the blueprint. Restricting a study to one narrow group removes a confounder by holding it fixed, and matching each exposed person to a similar unexposed one builds balance in from the start. Cleverer designs use a genetic variant fixed at conception as a stand-in for an exposure, since the genes you are born with cannot be confounded by habits you later adopt. That sits close to my genetic work, where the question is always whether a variant causes a trait or merely travels with the culprit.

What you can do as a reader

You do not need to run the statistics to read defensively. When you meet a claim that A affects B, ask whether the comparison came from a randomized trial or from observation, because that one fact changes how much trust the association has earned. If it was observational, ask what was adjusted for, and what could not be. A study that names its likely confounders and discusses the ones it could not rule out is being honest with you. One that reports a strong association and stops there is handing you a correlation and inviting you to supply the causation yourself.

The point is not to distrust every association, since some are genuinely causal. A real, repeatable link is the beginning of an investigation, and the whole discipline lies in asking what else was in the room.